HYPERSPECTRAL HYPERION IMAGERY ANALYSIS AND ITS APPLICATION USING SPECTRAL ANALYSIS

Rapid advancement in remote sensing open new avenues to explore the hyperspectral Hyperion imagery pre-processing techniques, analysis and application for land use mapping. The hyperspectral data consists of 242 bands out of which 196 calibrated/useful bands are available for hyperspectral applications. Atmospheric correction applied to the hyperspectral calibrated bands make the data more useful for its further processing/ application. Principal component (PC) analysis applied to the hyperspectral calibrated bands reduced the dimensionality of the data and it is found that 99% of the data is held in first 10 PCs. Feature extraction is one of the important application by using vegetation delineation and normalized difference vegetation index. The machine learning classifiers uses the technique to identify the pixels having significant difference in the spectral signature which is very useful for classification of an image. Supervised machine learning classifier technique has been used for classification of hyperspectral image which resulted in overall efficiency of 86.6703 and Kappa co-efficient of 0.7998.


INTRODUCTION
Hyperion instrument can capture 256 spectra each with 242 spectral bands.(Barry, 2001;Beck, 2003& Pengra, Bruce W, Johnston, Carol A, Loveland, Thomas R , 2007) .Hyperion covers the area perpendicular to the motion of the satellite (Kruse, F. A ,1996;Kruse, F. A., 2003;Kruse, FA, Boardman, 2002).Land cover thematic mapping can be determined using remote sensing data to provide important information for performing temporal land cover change analysis (Kavzoglu, 2009).Thus for thematic information extraction several previous studies employed multispectral imagery for land use/cover mapping application (Canty, in press ;Dixon, 2008;Huang, 2002;Nemmour,2006).In hyperspectral image, imaging system such as Thematic Mapper, Landsat Multi Spectral System or SPOT can be used for land surface cover features (Hong SY, 2002;Huete AR, 2003).Supervised classification can be used for classification and is defined as the process of using samples of known classes to classify the remaining unknown pixels to these classes with in the image (Campbell, J. B. 1996).In supervised classification, estimates are derived from the training samples which include number of classes be specified in advance (Plaza, 2005;Plaza, 2009;Small, C. 2001).Using Hyperion hyperspectral imagery, accuracy of different classification approaches for land use mapping is rare in the literature (Du, P.,2010;Pignatti, S., 2009;Walsh, S. J., 2008;Wang, J., 2010).
Since large number of bands is available in Hyperion hyperspectral image, therefore its pre-processing is different and is required before further analysis.The processed hyperspectral data can be used for different application after reducing the volume and dimensionality of the data.The processed hyperspectral data also enable traditional classification methods application on few selected bands having relevant information.In this paper pre-processing of Hyperion Hyperspectral orthoimage, application of Quick Atmospheric Correction , Principle Component analysis, vegetation delineation and normalized difference vegetation index, spectral profile of different classes and machine learning supervised classifier i.e spectral angle mapper will be used to achieve higher overall efficiency of classification.

STUDY AREA
The study area, is EO11500372005285110KF_1T.The Hyperion data are provided in GeoTIFF format.The Hyperion product includes a metadata file and multiple image bands.The product corner fields within the metadata files reflect the corners of the image area.

DATA SETS
The imagery is orthoimage and was acquired as a full long scene i.e 185-km tile and processing level 1 (L1_T).

DATA PRE-PROCESSING
Hyperion includes digital number to radiance transformation, radiance to reflectance conversion and atmospheric corrections /reflectance retrieval.

DN To Radiance Conversion
EO1-Hyperion hyperspectral image consists of number of continuous spectral bands, each pixel of which stored the energy as a digital number (DN).Stacked image is used to convert DN into Radiance values.The digital numbers were stored as 16-bit signed integer.Image was converted into absolute radiance by using following equation.Each band of NIR (1 to 70) and SWIR (71 to 242) was divided by its scale factor i.e 40 and 80 respectively (Thenkabail PS, 2004a).

VNIR , SWIR 40 80
The image is stored in ENVI Standard format and then it is converted in BIL (Bit in Line) data format.

Radiance to Reflectance Conversion
To convert the radiance into reflectance, following formula is used on individual band and was stacked in further processing steps (ThenkabailPS,2004b): (2)

Quick Atmospheric Correction (QUAC)
It is a scene based empirical approach used for the removal of atmospheric effects.It is based on the radiance values of the image/scene.QUAC model provides atmospheric correction of multispectral and hyperspectral imagery in VNIR to SWIR wavelength ranges.As compared to other methods, it used atmospheric compensation factors directly from the information contained within the image scene, without ancillary information.It has relatively faster computational speed as compared to other methods.QUAC provided better retrieval of reasonable reflectance spectra even if an image didn't have proper wavelength or radiometric calibration or solar illumination intensity be unknown (Agrawal, 2011).Preprocessing on the hyperspectral Hyperion orthoimage imagery was carried out by using the Hyperion tool.savtoolkit and was converted in ENVI into ENVI format files that contain information of bad band, wavelength, full width half maximum.Subsequently QUAC is applied in ENVI to provide atmospheric correction to hyperspectral imagery in VNIR to SWIR wavelength ranges.QUAC will provide better results for further processing.

RESULTS AND DISCUSSION
Hyperion orthoimage raw data analysis revealed that out of the total of 242 bands, 44 non calibrated bands have zero values which are set during level 1B pre-processing.Zero band values are bands from 1 to 7, bands from 58 to 76 and bands from 225 to 242.Resultantly, 198 bands were established to be useful for further analysis.

Principal Component Analysis
PCA was applied on the Atmospheric corrected data set of 155 bands of Hyperion orthoimage.As shown in The PC1 band contains the largest percentage of data variance and is highly uncorrelated.PC2 band contain the second largest and PC3 contain the third largest data variance and is also uncorrelated.PC4 to PC10 bands appear noisy as they contain very little variance.PC1, PC2 and PC3 can be used to produce more colourful colour composite images than spectral colour composite images because the data of PCs bands are highly uncorrelated.

Spectral Profile of Water and Buildup area
Water and build-up were also identified using the spectral characteristics of the features.Figure 4 and 5 shows spectral profile of water with radiance and reflectance respectively.Figure 6 and 7 shows the spectral profile of build-up area with radiance and reflectance respectively.

CONCLUSION
Hyperion sensor is today the only ''real'' space borne hyperspectral sensor in orbit, acquiring spectral information of Earth's surface objects in 242 spectral bands and at spatial resolution of 30 m. Out of 242 spectral bands, 155 calibrated bands are used for further processing.QUAC, quick atmospheric correction has been applied on 155 calibrated bands which have lowered the reflectance of the image in the blue and red region whereas it increases the value of reflectance in the NIR and SWIR region as compared to the apparent reflectance.Principal component analysis is applied to reduce the dimensionality and use the data as conventional bands.From the PCA, it is evident that first 10 PCs contributed more than 99 percent of the information.In this data set of 155 bands, 97.84 % data variability was explained by the first (PC).Another 3 PCs contributed 1.8% variability.Thus the dimensionality of the data is around four.NDVI and vegetation delineation is used to for vegetation classes feature extraction.Spectral profiles are used for feature extraction of water and build-up areas of the study area.Spectral angle mapper classification, uses an n-D angle to match pixels to reference spectra is a good approach for feature extraction.Accuracy assessment of the derived classification showed the overall efficiency of supervised machine learning classifier spectral angle mapper resulted in 86.6703 and Kappa co-efficient of 0.7998 on hyperspectral image.
less planetary reflectance Lλ= Spectral radiance at the sensor's aperture d = Earth-Sun distance in Astronomical units ESUN  = Mean solar Exo-atmospheric irradiances θs= Solar zenith angle in degrees Earth-sun distance was calculated using following equation d= 1-0.01672*Cos (0.9856*(Julian Day-4)) Figure1.Principal Component (PC) image display (a) PC1 (b) PC2 (c) PC3 (d) PC4 (e) PC5 (f) PC6 (g) PC7 (h) PC8(i) PC9 (j) PC105.2NDVI (Normalized Difference Vegetation Index) and Vegetation DelineationThe surface was covered with many different features which include rocks, vegetation cover, water body and roads.Large area was covered with vegetation class.It was necessary to mask out the vegetation areas.For this purpose NDVI was calculated and vegetation was delineated.Generalized formula for NDVI is as follows: NDVI = NIR-Red (3) NIR+ Red

Figure 4 .
Figure 4. Spectral Profile of Water with radiance

Figure 8 .
Figure 8.The Acquired Hyperion Orthoimage (top image) and Subset of Acquired Hyperion Orthoimage (bottom image) covering the studied area5.5 Classification Accuracy AssessmentClassification accuracy assessment was developed and implemented in ENVI based on the confusion matrix analysis of the maps produced from the implementation of the spectral angle mapper classification technique on the Hyperion orthoimage.Thus overall accuracy and Kappa coefficient were calculated.

Table 1 ,
First 10 PCs contain more than 99 percent of the information in a data set of 155 bands.First (PC) contain 97.84 percent of the information.Second, third and fourth PCs contain 1.8 percent of the information.Thus it lead to the conclusion that the dimensionality of the data is around four.